Research And Application Of Medical Image Segmentation Method Based On Convolutional Neural Network

Accurate identification and segmentation of target tumors or organs from medical images such as B-mode ultrasound and CT has always been one of the key issues in medical image analysis.At present,lung CT scan images are mainly used in lung cancer diagnosis to manually screen and identify lung cancer.The lung CT data is increasing exponentially,which brings huge challenges and burdens to physicians’ manual screening work.Therefore,we mainly conduct research based on lung cancer data.Using deep learning methods for lung cancer image segmentation can greatly improve the diagnosis efficiency and accuracy of physicians.It can assist clinical diagnosis and virtual surgery.However,the existing medical image segmentation methods based on convolutional neural networks still have problems such as loss of edge details and poor boundary segmentation accuracy under complex backgrounds.In addition,lung cancer patients need to determine their subtypes through biopsy,which is harmful to patients.However,due to the current lack of public training data sets and powerful artificial intelligence models,it is still a challenge to identify the pathological types of early lung cancer through CT images.In response to the above problems,this thesis is based on generative adversarial networks and convolutional neural networks and carries out related research on image segmentation algorithms and lung cancer recognition.The main contents include:1.In view of the unrefined detection and segmentation boundary of missing small targets in existing medical image segmentation,a new segmentation adversarial network(SA-UNet)based on U-Net and generated adversarial architecture was proposed for lung cancer CT image segmentation.Based on the U-Net architecture,the segmentation network module adds three modules: residual module,atrous convolution module and feature pyramid pooling module to extract semantic features at multiple scales and strengthen medicine image feature extraction capability.The discrimination network module is used to determine whether the generated medical image segmentation map or the real medical image segmentation map is generated.At the same time,it sets the anti-loss and improves edge segmentation in complex backgrounds accuracy makes it more and more close to the real image.The experimental results show that the SA-UNet method is superior to the original FCN,Seg Net,U-Net and other latest methods in lung cancer CT image segmentation.In this thesis,SA-UNet is trained and tested on the Data Science Bowl 2018 public data set,showing that the model has good generalization.2.Based on the segmentation adversarial network,a deep convolutional neural network SA-VGG16 is proposed for lung cancer subtype identification.First,the segmentation adversarial network is used to preprocess the lung cancer subtype data set and to enhance the data.Secondly,the processed image is used as the input data of the lung cancer subtype recognition algorithm to train the lung cancer subtype recognition algorithm.Then,the boosting strategy is used to train multiple SA-VGG16 as weak classifiers to solve the problem of imbalance of original category data.This is the first attempt to use a deep model to identify the pathological type of lung cancer early from small-scale CT images.Experimental results show that the performance of this algorithm in identifying pathological types of lung cancer has been greatly improved.